In ordinary crystals there is a clean rule: if a band of electrons is topological, meaning it carries a nonzero Chern number (an integer that counts how the quantum states twist across momentum space), then its states cannot be repackaged into a set of well-localized, orthonormal real-space orbitals called Wannier functions. Stubbs (2024) proved a version of this rule for materials with no periodicity at all, such as alloys, glasses and quasicrystals. If the occupied states admit an orthonormal basis whose orbitals decay faster than $1/r^{2+\delta}$ for some $\delta>0$, then the real-space Chern marker vanishes. The result is stated for infinite systems, and the research file records that the paper contains no numerical experiments of any kind. Nobody knows how fast the theorem "turns on" in a finite sample, whether the decay threshold $1/r^2$ is sharp in practice, or whether the box-truncation device the proof relies on is a usable numerical diagnostic or only a proof tool.
We propose the first quantitative empirical test of this theorem. It has three parts. The first is a construction that produces orthonormal bases of the same occupied subspace with a decay exponent we choose. The second measures the truncated Chern marker the proof uses, as a function of box size, decay exponent and disorder strength. The third asks whether bases found in topological phases ever beat the $1/r^2$ threshold. The concrete claim is that the theorem's infinite-volume statement has a measurable finite-size form, with a convergence rate controlled by $\delta$, and that the borderline $s=1$, which the proof leaves open, can be probed directly by numerics. The deliverable is a set of scaling laws plus a public disorder ensemble on which these real-space claims can be checked.
Core Claim
Write an orbital in a generalized Wannier basis as $s$-localized if its amplitude decays like $|x|^{-2s}$. Theorem 1 of Stubbs (2024) assumes $s=1+\delta$. The proof replaces the spatial cutoff in the Chern marker with a finite-rank projector $P_L$, built from basis orbitals whose centers lie in a box of side $L$. It then shows that the Hilbert–Schmidt tails of $(P-P_L)XP_L$ and related commutators vanish as $L\to\infty$.
We conjecture that the truncated marker obeys
$$|C_L(P)| \le K\, L^{-\gamma(\delta)},$$
where $C_L(P)$ is the Chern marker evaluated with $P_L$ in place of the cutoff, $K$ is a sample-dependent constant, and $\gamma(\delta)>0$ is a rate inherited from those tail bounds. We expect $\gamma$ to grow with $\delta$ because the bounds are integrated power-law tails. We will extract the rate implied by Propositions 2.4–2.5 analytically and compare it to fitted values.
The second half of the claim is falsifiable. In topological phases, every orthonormal basis we can construct should have fitted exponent $s\le 1$ within error bars. As disorder drives the system toward the topological transition, the best achievable $s$ should approach 1 from below rather than jump. A clean violation in either direction would be informative. It would mean either that finite systems mislead badly, or that the borderline behaves differently from what the theorem's hypotheses suggest.
Research Program
Tunable-decay bases. The method comes from Stubbs et al. (2022) and Stubbs et al. (2021). Their iterated projected position (IPP) algorithm diagonalizes the position operator restricted to the occupied subspace, $PXP$. It splits the spectrum at gaps into sub-bands, then diagonalizes $Y$ within each sub-band. When those spectral gaps exist uniformly, the result is exponentially localized generalized Wannier functions.
In trivial phases we start from the IPP basis. We then rotate it by a unitary $U=e^{iH_s}$, where $H_s$ is a random Hermitian matrix on the Wannier centers with entries decaying as $|x-y|^{-2s}$. In two dimensions, power-law-decaying matrices with exponent above the dimension form an algebra, so $U$ inherits the prescribed decay. The rotated basis spans exactly the same projector $P$, whose true Chern marker is zero, and its tails are decaying at a rate we set. We then sweep $s$ across 0.6 to 2.
Two outcomes are possible. If $C_L$ fails to converge to zero for $s<1$, even though $C(P)=0$, then the $P_L$ substitution is a proof device that needs the decay hypothesis. If $C_L$ converges anyway, then the hypothesis is stronger than necessary for this route.
Best achievable decay in topological phases. For Chern phases we cannot inject localization, so instead we search for it. We use three basis constructors: IPP itself, which degrades when $PXP$ gaps close; real-space spread minimization in a large supercell; and IPP applied to a slightly perturbed projector followed by projection back. On each basis we fit decay exponents on log–log tails.
Spectral gap statistics. We measure the distribution of $PXP$ gap sizes across disorder realizations. This addresses an open question in Stubbs et al. (2022): whether the uniform-gap assumption is equivalent to triviality. Our version is empirical, asking whether the gap and the marker carry the same information sample by sample.
Topologically localized insulators. Lapierre et al. (2022) showed that three-dimensional insulators can have every bulk eigenstate exponentially localized yet a nonzero bulk invariant $N_{TLI}$. The occupied eigenstates are then themselves a localized basis. This system tests which hypothesis carries the theorem. Stubbs assumes a spectral gap with an exponentially localized projector, whereas the TLI has only a mobility gap: its states are localized, but there are no energy gaps between them. Stubbs also works with a two-dimensional Chern marker, whereas the TLI invariant is genuinely three-dimensional. We will compute layer-resolved Chern markers, the decay of the eigenstate basis, and IPP spectra inside the TLI bulk below the metallic threshold $\lambda_c\approx 0.088$ reported there.
Baselines
The naive reference point is the standard local Chern marker computed with a sharp spatial cutoff on the same samples, ignoring any basis. It shows what one learns about topology without the theorem's machinery. It runs on identical disorder realizations and box sizes, and it costs one projector computation per sample, which we match. Our $P_L$ marker beats it only if it shows a cleaner or faster approach to the asymptotic value at equal $L$, or if it separates the $s>1$ and $s<1$ regimes where the cutoff marker cannot. We do not expect it to beat the cutoff marker as a topology detector, and we will report that honestly.
The strongest existing method is IPP gap closure as a real-space topological signature, from Stubbs et al. (2022). That paper shows the closure qualitatively, in eigenvalue plots on the Haldane model, and reports no quantitative accuracy; no reported result found on Lacuna for classification accuracy of PXP gap closure on disorder ensembles. We will run IPP on the same lattices, with the same exponential position operators for periodic boundaries and the same diagonalization budget. We score how well a gap-closure threshold predicts the sign of the converged marker, alongside how well our fitted exponent does. Agreement is the expected result. Disagreement on a subset of samples near the transition, where one diagnostic stays correct, would be the interesting outcome.
For machine-learned topology prediction, the only numbers in the file are momentum-space ones. On Haldane Bundles, a Vision Transformer reaches 29.1% accuracy against a 21.0% chance level. GEBLNet (Gauge Equivariant Neural Networks) reaches 95.9% at 4 filled bands and drops to 52.5% at 8. Neither applies to disordered real-space projectors, so we treat these numbers as context rather than as a direct competitor.
The key ablation isolates the controlled-decay ingredient. We compare $C_L$ computed from the rotated bases to $C_L$ computed from the unrotated exponential IPP basis of the same projector. Any difference in convergence is then attributable only to tail decay.
Why not just…?
Why not just compute the local Chern marker on bigger samples? For deciding whether a given sample is topological, this would plausibly work as well, and we do not claim otherwise. The marker is a measurement in our study, not a competitor. What it cannot tell you is how localization quality governs finite-size convergence, or whether the $1/r^2$ threshold is sharp. Those are the questions the theorem raises and leaves unmeasured. The trivial-phase rotation experiment answers them, and the marker alone cannot.
Why not just use Marzari–Vanderbilt spread minimization? Stubbs et al. (2021) report that spread minimization needs a careful initial guess, gets trapped in local minima, and leans on Bloch structure. In a large supercell it is still usable as a real-space spread minimizer, and we include it as one of the three basis constructors. If it finds bases with $s>1$ in a Chern phase, that would directly challenge our second claim, and the experiment is designed to catch it.
Why not just use level-spacing statistics to certify triviality? Lapierre et al. (2022) refute this. Their TLI has Poisson level statistics, meaning fully localized states, together with a nonzero $N_{TLI}$. That is exactly why we include the TLI. It is the regime where localization of eigenstates and triviality come apart, and we can check which ingredient of Stubbs' theorem the TLI evades.
Why not train a neural network on real-space projectors? The evidence in Haldane Bundles is that generic networks fall to near chance once local shortcuts are removed. GEBLNet (Gauge Equivariant Neural Networks) requires a Brillouin zone. More basically, a classifier would not test the theorem's quantitative content even if it worked.
Proposed Experiments
The ensembles are the disordered Haldane model on honeycomb lattices with onsite disorder strength $W$ swept across the transition, the Kane–Mele model with the time-reversal breaker of Stubbs et al. (2021), the $p_x+ip_y$ model on the Ammann–Beenker quasicrystal, and the stacked-Chern TLI model of Lapierre et al. (2022). Linear sizes will run up to the largest $L$ where dense diagonalization remains feasible. We will record wall-clock time and memory for IPP throughout, since the file notes that its cost has never been measured.
The first decisive check is the trivial-phase sweep. We plot $|C_L|$ against $L$ for each injected $s$ and fit $\gamma(\delta)$. Clean power-law convergence for $s>1$, with $\gamma$ increasing in $\delta$, confirms the finite-size form. We will also see whether convergence breaks down below $s=1$.
The second check is the topological-phase search. We look at the maximum fitted $s$ over all constructors and $10^2$ disorder realizations per $W$. A consistent ceiling at or below 1, approached continuously near the transition, supports the dichotomy at the borderline. Any robust exceedance falsifies it.
The third check compares gap statistics with the marker sample by sample. Agreement is evidence for the gap–triviality equivalence conjectured in Stubbs et al. (2022); a population of gapless-yet-trivial samples is evidence against it.
The fourth check is the TLI bulk analysis below $\lambda_c\approx0.088$. Nonzero layer-resolved two-dimensional markers alongside an exponentially localized eigenstate basis would mean the spectral-gap hypothesis, not the decay hypothesis, is what the theorem truly needs. Vanishing layer markers would place the TLI invariant entirely outside the theorem's scope.
References
- Algebraic Localization of Wannier Functions Implies Chern Triviality in Non-periodic Insulators
- Existence and Computation of Generalized Wannier Functions for Non-Periodic Systems in Two Dimensions and Higher
- Iterated projected position algorithm for constructing exponentially localized generalized Wannier functions for periodic and nonperiodic insulators in two dimensions and higher
- Topologically Localized Insulators
- Haldane Bundles: A Dataset for Learning to Predict the Chern Number of Line Bundles on the Torus
- Learning Chern Numbers of Multiband Topological Insulators with Gauge Equivariant Neural Networks